Development of Pneumonia Identification System: A Comparative Analysis of Some Selected CNN Architectures Using Adam, Nadam, and RAdam Optimizers

The most common cause of disease and mortality in Sub-Saharan Africa is pneumonia, a lung infection brought on by fungus, bacteria, and other pathogens. Health facilities in the region have limited radiologists coupled with the slow and subjective traditional diagnosis using chest X-rays. Deep learning methods such as convolutional neural networks, have proven useful in the medical diagnostic process. In this study, models were trained with six different pre-trained CNN architectures: MobileNetV2, VGG16, InceptionResNetV2, ResNet152V2, InceptionV3 and DenseNet201, as well as a custom CNN architecture with the aim of developing a user-friendly mobile application for pneumonia diagnosis. The influence of data augmentation on model performance was also investigated. The performance of the selected CNN models was evaluated using the Adam, Nadam, and RAdam optimizers. The top-performing model was the custom CNN model, which attained an accuracy score of 97.09% and an F1-score of 98.01% while employing the Nadam optimizer with no data augmentation. MobileNetV2 and ResNet152V2 both earned an accuracy score of around 96% while using the Nadam and the Adam optimizers respectively. Among all the models, the VGG16 model, which was trained using the Nadam optimizer with the application of data augmentation, achieved the lowest accuracy score. ResNet152V2 performed creditably well across all the scenarios and therefore the mobile application was developed using the ResNet152V2 model.